thanks for the idea for white matter 'only' lesion segmentation using a
This step may actually help in creating a white matter only mask which after
coregistering on the images of interest could reduce the numbers of classes
you have and ease the segmentation process.
I found that applying FAST on FLAIR images actually detects hyperintense
lesions very sensitively when you use 5 or 6 classes. In combination with a
thresholding step (e. g. >0.1) and a 'cleaning up step' that deletes very
small lesions (e. g. < 10 cont. voxels) it gives a pretty good result, about
in agreement with what a radiologist would classify or 'count' as lesion.
However, hyperintense rims remain at the peripheral gray matter and these
could be eliminated by the step proposed by you (as long as severe lesions
appearing as black holes do not disturb the idenitification of white matter
too much, but with heavily weighted priors it may work).